THE REPOSITORY'S OWN DOCUMENT · OPERANT

AD-006 · Bodies are fenced from the supervisor

A body cannot read the supervisor environment, cannot inherit chair authority, and dies with its slot - the isolation boundary as a decision.

Repository
agentic-harness/operant/docs/decisions
Commit
fbe90019
Date
2026-11-01
Length
46 lines · 3 min

Read the source on GitHub This page renders that file; it is not a summary of it.


ID: "AD-006" Title: "Holon bodies are fenced from the supervisor's secrets and their slot" Status: "Active" Date: "2026-10-19" Description: "A Holon body cannot read the supervisor's environment, cannot inherit chair authority, and dies with its slot"

AD-006: Holon bodies are fenced from the supervisor's secrets and their slot

Immutable once numbered: a correction is a new record that supersedes this one.

Context

A body is a same-uid child of the supervisor. The Holon red-team found five paths out of its slot:

  • OS-1: the body reads /proc/$PPID/environ, which is the supervisor's environment (channel tokens, provider keys).
  • OS-2: HOME is in the body's baseline, so ~/.config/holon/keys/<instance>.chair is reachable. Holon's principal() falls back to that file when no run token is set, so a body that unsets HOLON_RUN_TOKEN can run holon run finish as chair.
  • OS-6: Ctrl-C or SIGTERM on supervise killed the supervisor and not its body's process group. The body ran until its own timeout while its lease expired and the kernel re-claimed the stage, so the same stage could run twice.
  • OS-7: every non-running run status was treated as a clean self-submit.
  • OS-10: the timeout kill sent SIGTERM to the group and SIGKILL only to the leader, so a grandchild that traps TERM survived.

Decision

  1. The supervisor clears PR_SET_DUMPABLE before every body spawn (Linux). exec resets the flag, so each body stays dumpable; only the supervisor's /proc entries close.
  2. Every body environment pins HOLON_CHAIR_KEY to a constant that no chair digest matches. Holon reads that override before the key file, so the CLI's implicit chair lookup resolves to anonymous.
  3. A body's process group is SIGKILLed when the body's future is dropped. supervise races its whole loop against SIGINT, SIGTERM and SIGHUP, so a signalled supervisor takes its body with it at once instead of waiting for lease expiry.
  4. The run's kernel status decides the slot (closed_by). succeeded and failed are a self-submit and a clean stop. running is open. Any other status (timeout, abandoned, or one this code does not know) is a reclaim: the body is killed and finish is skipped, because the kernel already closed the run.
  5. A kill of a body's group is SIGTERM, then a 5 s grace (KILL_GRACE), then SIGKILL to the group.

Consequences

  • Residual, not closed here: a same-uid body can still cat the operator-owned chair key file explicitly. Closing that needs act bodies under a separate uid, which is the O-11 isolation work owned by the Rust runner. Until it lands, an act body is trusted with the operator's files.
  • Off Linux there is no dumpable flag. The supervisor logs a warning for each body it spawns, and the environment stays readable there.
  • A signalled supervisor kills its body with SIGKILL, so the body loses in-flight work. The run then closes by lease expiry as timeout.
  • A SIGKILLed supervisor cannot run any of this. Its body keeps running until the body's own timeout, and the kernel may re-claim the stage in the meantime. This decision does not cover that case.
  • Heartbeat failure still only warns. The status poll (every 15 s) detects a reclaim, so one failed heartbeat does not kill a live body.

Orthogonality

AD-007 covers the fail-closed policy check, and AD-008 covers the Holon-side notify and driver fixes. The observe-posture denial of holon_skill_invoke (OS-4) was decided in iter-963 and is not changed here.

Revisions

  • 2026-10-19: created

Other published documents in operant

Back to the operant case study 23 of this study's documents are published on the site; the repository holds the rest.